(458h) Online Feedback Control of an Atomic Layer Etching Process Using Model Predictive Control Based on a Transformer Model

As the downscaling of semiconductor feature sizes continues, there have been numerous obstacles to the fabrication of these devices in the nanoscale. Atomic layer etching (ALE) is a component of semiconductor processing in which thin monolayers of high-κ oxide films are etched on transistor surfaces, which must exhibit self-limiting and high surface uniformity characteristics. Despite the employing of ALE, the semiconductor fabrication industry has faced hurdles with maintaining product conformance and identifying causes for defects, due to the limitations in the responsiveness of control systems to detect and correct disturbances [1]. Additionally, there is a lack of meaningful industrial data for these processes from the experimental scale to allow these processes to integrate into industrial practices. Thus, in silico simulation has provided a substitute for the generation of data that will improve the performance (e.g., responsiveness and correction) of online feedback controllers.

First, a multiscale computational fluid dynamics (CFD) framework is proposed for a previously developed ALE process for Al₂O₃ films [2] that conjoins a mesoscopic kinetic Monte Carlo (kMC) simulation with a macroscopic CFD model. Next, a collection of time-series data for multiple inputs (e.g., flow rate) and a single output (e.g., etching per cycle or EPC) is extracted to construct a predictive model. This predictive model is established using a transformer, which is efficient at natural language processing and is recognized for identifying relationships between sequences of aggregated data sets [3]. Following the development of a transformer model, an online feedback controller, a proportional-integral (PI) controller that is tuned appropriately, will be integrated to the multiscale CFD simulation, which is purposefully perturbed by various disturbances (e.g., shift and drift), to correct multiple input parameters and bring the process to a user-defined set-point. With the importance of optimizing the correction made to the input parameters, this work will also compare the performance of a PI controller with that of a model predictive controller (MPC), which is desired for minimizing production costs (e.g., reagent consumption) [4].

• Butler, S. W., 1995. Process control in semiconductor manufacturing. Journal of Vacuum Science & Technology B: Microelectronics and Nanometer Structures Processing, Measurement, and
  Phenomena 13, 1917–1923.
• Yun, S., Tom, M., Ou, F., Orkoulas, G., Christofides, P. D., 2022. Multiscale computational fluid dynamics modeling of thermal atomic layer etching: Application to chamber configuration design. Computers & Chemical Engineering, 161, 107757.
• Zhang, C., Yella, J., Huang, Y., Qian, X., Petrov, S., Rzhetsky, A., Bom, S., 2021. Soft sensing transformer: Hundreds of sensors are worth a single word. In: 2021 IEEE International Conference on Big Data, 1999–2008, Orlando, FL, USA.
• Xiao, T., Wu, Z., Christofides, P. D., Armaou, A., Ni, D., 2022. Recurrent neural-network-based model predictive control of a plasma etch process, Industrial & Engineering Chemistry Research, 61, 638–